Don't Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, Anthropic

What Are Agent Skills and Why Are They Important?

Introduction to Agents and Skills

• The speakers, Barry and Mahes, discuss the evolution of agents from their previous talk, highlighting that many now use agents daily but still notice gaps in expertise.

• They explain the shift from building agents to focusing on skills due to advancements like MCP becoming a standard for agent connectivity.

New Paradigm for Agents

• A new paradigm is emerging where there is a tighter coupling between models and runtime environments; they argue that code serves as a universal interface to the digital world.

• The realization that while customization is important, the underlying agent structure can be more universal than previously thought.

Domain Expertise vs. General Intelligence

• The speakers emphasize the importance of domain expertise by comparing an intelligent individual (Mahes) with an experienced professional (Barry), suggesting that expertise leads to better outcomes.

• Current agents are likened to Mahes—intelligent but lacking necessary contextual knowledge and learning capabilities over time.

Concept of Agent Skills

• Agent skills are defined as organized collections of files containing procedural knowledge designed for easy use by both humans and agents.

• This simplicity allows anyone with a computer to create and utilize these skills, integrating seamlessly with existing tools like Git or Google Drive.

Advantages of Using Code in Skills

• Traditional tools often have ambiguous instructions; however, code is self-documenting, modifiable, and can reside in file systems until needed.

• An example illustrates how repetitive tasks can be streamlined by saving scripts within skills for future use, enhancing consistency and efficiency.

The Structure and Growth of Skills

Organization of Skills

• Skills are progressively disclosed at runtime; only essential metadata is shown initially while detailed instructions remain accessible when needed.

Types of Skills Developed

• Since launching five weeks ago, thousands of skills have emerged across various categories: foundational skills, third-party partner skills, and enterprise-specific skills.

Foundational Skills

• Foundational skills provide new general or domain-specific capabilities; examples include document creation/editing abilities added to Claude's functionality.

Third-party Partner Contributions

• Partners have developed specific skills tailored for their software products; Browserbase created a skill for browser automation enhancing Claude's web navigation capabilities.

Enterprise-Specific Applications

• Large enterprises are leveraging these skills as educational tools within organizations; discussions with Fortune 100 companies reveal significant interest in this approach.

Understanding the Evolution of Skills in Agent Technology

The Role of Skills in Developer Productivity

• Discussion on how organizations utilize bespoke internal software to enhance developer productivity, focusing on large teams serving thousands of developers.

• Emphasis on the creation and consumption of skills by various individuals within an organization, highlighting that anyone can create these skills to enhance agent capabilities.

Trends in Skill Complexity

• Observation that skills are becoming increasingly complex, evolving from simple markdown files to comprehensive packages including software and executables.

• Noting that while basic skills may take minutes to develop, more advanced skills could require weeks or months for proper construction and maintenance.

Integration with MCP Servers

• Explanation of how developers are building skills that orchestrate workflows across multiple MCP tools, enhancing functionality with external data connectivity.

• Exciting trend where non-technical individuals (e.g., finance, legal professionals) are creating skills, validating the idea that coding is not a prerequisite for skill development.

Emerging Architecture of General Agents

• Introduction to the concept of an agent loop managing internal context and token flow, coupled with a runtime environment for file system access.

• Description of how agents can be equipped with libraries of hundreds or thousands of skills relevant at runtime based on specific tasks.

Future Directions for Skills Development

• Insights into Enthropic's deployment strategy using Claude across new verticals like financial services and life sciences through tailored MCP servers and skill sets.

• Focus on treating skills similarly to software by exploring testing methodologies, versioning practices, and quality measurement tools as they evolve over time.

The Vision for Knowledge Sharing

• Discussion about the potential value derived from sharing and distributing skills within organizations as a collective knowledge base evolves.

• Highlighting how agents improve through interaction and feedback mechanisms, ensuring they adapt to team-specific needs over time.

Understanding the Evolution of Skills in AI

The Growing Ecosystem of Skills

• The development of skills within the AI ecosystem enhances the utility of agents, extending their value beyond individual organizations to the broader community.

• As skills evolve, they contribute to a more reliable and useful knowledge base, facilitating continuous learning for AI systems.

Standardized Learning and Memory

• Claude's standardized format ensures that any information recorded can be efficiently utilized by future versions, making learning transferable.

• Skills allow Claude to acquire new capabilities quickly and discard obsolete ones, enhancing cost-effectiveness for frequently changing information.

Continuous Improvement Over Time

• The goal is for Claude to significantly improve from day one to day thirty of interaction, showcasing its evolving capabilities through skill creation.

Comparing Agent Stack with Computing Models

• The analogy between models as processors highlights their potential when combined with an operating system that orchestrates resources effectively.

• Agent runtime is likened to an operating system that maximizes processor value by managing data flow and processes around it.

Opening Up Creativity Through Skills

• While few companies create processors or operating systems, millions develop applications that encode unique expertise; skills aim to democratize this layer.

• The focus is shifting from rebuilding agents to building skills as a new paradigm for sharing capabilities, inviting collaboration in skill development.